IsaacLab Training#

Reusable downstream Isaac Lab training repository for custom task formulations, shared MDP terms, and policy export workflows.

Note

To train a simulation policy for your specific robot, first generate a calibrated robot USD from its kinematics, as described in Generate a Calibrated USD for Isaac Lab Training. Then pass it to training with env.scene.robot.spawn.usd_path.

Get Started#

  1. Setup: install the training package with Isaac Sim and run a short training test.

  2. Training a DisplayPort Cable Insertion Policy and ROS Deployment: train the DisplayPort insertion policy, export it with LEAPP, and deploy it with Isaac ROS. It also lists training times and recommended settings.

Trainable Policies#

This repository currently packages DisplayPort cable insertion policies for the Flexiv Rizon 4s arm with a Grav parallel gripper.

Policy

Task ID

Notes

DisplayPort task-space insertion (PhysX)

IsaacTraining-DisplayPortInsertion-Rizon4s-TaskSpace-ROS-Inference

Recommended. More stable training and the best results in real-robot deployment.

DisplayPort task-space insertion (Newton)

IsaacTraining-DisplayPortInsertion-Rizon4s-TaskSpace-Newton-ROS-Inference

Experimental. Newton 1.6 MJWarp profile; nominal by default with a calibrated-USD override.

Note

PhysX training is more stable, while Newton training is still experimental. We recommend the PhysX-trained policy for the best results in real-robot deployment.

For visualization and evaluation, play a training task id directly: the play scripts apply the configuration’s play_mode() overrides (fewer environments, wider spacing, observation corruption disabled). Pass --train_env_cfg to play the training configuration unchanged. Use the plain non-ROS-inference tasks for ablations you do not intend to deploy directly.

Repository Architecture#

The repository separates reusable Isaac Lab infrastructure from concrete task formulations.

Path

Purpose

src/isaaclab_training/mdp/

Shared actions, observations, rewards, events, terminations, and noise models.

src/isaaclab_training/tasks/

Registered task formulations. DisplayPort insertion is the first task.

src/isaaclab_training/cli/

Policy export and policy I/O validation commands.

src/isaaclab_training/export/

LEAPP export helpers used by the generic exporter.

tests/

Lightweight checks for task registration and export-specific tensor semantics.